AMD Radeon RX Vega to disrupt gaming, pro graphics, and AI

[HEXUS]

And AMD announces a "true engineering collaboration" with Bethesda.

Read more.

[scaryjim]

Hmmm ... RX Vega to go with RY Zen....?

[rob4001]

So how many cards on is just 1??

[CAPTAIN_ALLCAPS]

Well, another pointless event for a card still months from release.

Meanwhile GV100 has been spotted in drivers, implying that nVidia has the first Volta engineering samples back from the fabs.

[Nobull]

Embarrassing how long this is taking now. Nvidia look like lapping them a generation. Terrible news all round.

[CAT-THE-FIFTH]

Vulkan as standard in Bethesda games is a big deal - it really is!! Games like Skyrim and FO4 would benefit massively from Vulkan and I really hope the next Elder Scrolls game uses it.

Well, another pointless event for a card still months from release.

Meanwhile GV100 has been spotted in drivers, implying that nVidia has the first Volta engineering samples back from the fabs.

Embarrassing how long this is taking now. Nvidia look like lapping them a generation. Terrible news all round.

Maybe you should both stop reacting and start actually looking at some of the tech which they talked about. Pascal and Polaris are rejigged versions of previous designs. Pascal is an evolution of Maxwell,and Polaris is an evolution of the previous GCN1.2 Fiji and Tonga.

Vega is not a small redesign - its a HUGE redesign incorporating new features. Vega is the competitor to Volta.

Look at HBCC - during the demo they showed how they could cut down massively on VRAM yet get the same results - the rapid packed math demo showed them doubling GPU physics calculations.

Then on top of that they talked about the NCU - the compute units in Vega are redesigned over the previous AMD cards. Apparently the way tessellation is processed is also altered.

Also even rasterisation is handled in a new way,ie,similar to the tile based method Nvidia introduced with Maxwell.

They also said the NCU design was already used in consoles - this means the PS4 PRO design was a halfway point between Polaris and Vega.

Its the biggest redesign since GCN1.0 was released.

[CAT-THE-FIFTH]

Also something else Raja Koduri might have hinted at - one of the Vega GPUs have a similar amount of compute units as Fiji,ie,64.

[will19565]

Vulkan as standard in Bethesda games is a big deal - it really is!! Games like Skyrim and FO4 would benefit massively from Vulkan and I really hope the next Elder Scrolls game uses it.

Well, another pointless event for a card still months from release.

Meanwhile GV100 has been spotted in drivers, implying that nVidia has the first Volta engineering samples back from the fabs.

Embarrassing how long this is taking now. Nvidia look like lapping them a generation. Terrible news all round.

Maybe you should both stop reacting and start actually looking at some of the tech which they talked about. Pascal and Polaris are rejigged versions of previous designs. Pascal is an evolution of Maxwell,and Polaris is an evolution of the previous GCN1.2 Fiji and Tonga.

Vega is not a small redesign - its a HUGE redesign incorporating new features. Vega is the competitor to Volta.

Look at HBCC - during the demo they showed how they could cut down massively on VRAM yet get the same results - the rapid packed math demo showed them doubling GPU physics calculations.

Then on top of that they talked about the NCU - the compute units in Vega are redesigned over the previous AMD cards. Apparently the way tessellation is processed is also altered.

Also even rasterisation is handled in a new way,ie,similar to the tile based method Nvidia introduced with Maxwell.

They also said the NCU design was already used in consoles - this means the PS4 PRO design was a halfway point between Polaris and Vega.

Its the biggest redesign since GCN1.0 was released.

thats a nice idea, but what is really inportant is who will get their top GPU's out first.

[CAPTAIN_ALLCAPS]

Vulkan as standard in Bethesda games is a big deal - it really is!! Games like Skyrim and FO4 would benefit massively from Vulkan and I really hope the next Elder Scrolls game uses it.

Well, another pointless event for a card still months from release.

Meanwhile GV100 has been spotted in drivers, implying that nVidia has the first Volta engineering samples back from the fabs.

Embarrassing how long this is taking now. Nvidia look like lapping them a generation. Terrible news all round.

Maybe you should both stop reacting and start actually looking at some of the tech which they talked about. Pascal and Polaris are rejigged versions of previous designs. Pascal is an evolution of Maxwell,and Polaris is an evolution of the previous GCN1.2 Fiji and Tonga.

Vega is not a small redesign - its a HUGE redesign incorporating new features. Vega is the competitor to Volta.

Look at HBCC - during the demo they showed how they could cut down massively on VRAM yet get the same results - the rapid packed math demo showed them doubling GPU physics calculations.

Then on top of that they talked about the NCU - the compute units in Vega are redesigned over the previous AMD cards. Apparently the way tessellation is processed is also altered.

Also even rasterisation is handled in a new way,ie,similar to the tile based method Nvidia introduced with Maxwell.

They also said the NCU design was already used in consoles - this means the PS4 PRO design was a halfway point between Polaris and Vega.

Its the biggest redesign since GCN1.0 was released.

Talk is cheap, and marketing bumf is cheaper. Furthermore, many of these features sound like they need explicit support by the game developer - and I have learned to take a dim view of technologies like that and stick to real-world performance (and the only real-world performance data we have, Doom on Vulkan, isn't terribly impressive so there had better be large amounts of untapped potential).

If AMD intended for Vega to be more Volta than Pascal they shouldn't have limited their largest Polaris chip to 232mm2, leading to 1060-level performance at best and grotesquely overpriced top end chips from nVidia for essentially an entire generation.

[CAT-THE-FIFTH]

Talk is cheap, and marketing bumf is cheaper. Furthermore, many of these features sound like they need explicit support by the game developer - and I have learned to take a dim view of technologies like that and stick to real-world performance (and the only real-world performance data we have, Doom on Vulkan, isn't terribly impressive so there had better be large amounts of untapped potential).

If AMD intended for Vega to be more Volta than Pascal they shouldn't have limited their largest Polaris chip to 232mm2, leading to 1060-level performance at best and grotesquely overpriced top end chips from nVidia for essentially an entire generation.

Trolling is cheap - your attitude in both AMD CPU and GPU threads,you seem to want to twist things as negatively as possible. Yet I can see you hyping Volta already as the next best thing - talk is cheap and we had Nvidia focus group member Rollo who started a new account before a major AMD launch, to keep pushing that Nvidia was ready to launch their latest card when eventually it didn't come for months.

Funny how just before Ryzen and Vega which will probably be launching within a few months of each other we have a new account suddenly trying to push anything AMD in the most negative way. I have been here long enough to recognise it.

But you are scared that AMD will get Vega out before Volta since your E-PEEN is threatened and you know very well if Nvidia is only launching the GTX1080TI in the next week or two,Nvidia won't be releasing a more powerful card until the end of the year,but don't worry I expect when AMD does release Vega you will make cheap talk and keep on pushing that Volta is out soon.

[CAT-THE-FIFTH]

Plus some of you are in some bubble of reality about AMD not having a proper high end card - AMD on record said they diverted most of their R and D to Ryzen development in the last two years. Fiji itself was an HBM test vehicle - Tonga was probably partially paid for by Apple,since they got the first chips.

Its blatantly obvious to anybody they skipped a high-end Polaris chip due to cost and the fact they had issues scaling up with Fiji for gaming,which Polaris probably would not have fixed.

Polaris is internally identified as a GCN1.2 design in the optimisation guides.

This is what they have done - redesigned GCN so that it actually can scale up more. This is why some of you need to actually read what the changes are trying to do - increase efficiency and utilisation of the CUs which is the issue they have been having. Its all there,but why bother trying to understand what any of it is - easy to not bother.

The company is bleeding money on the CPU side - Ryzen is far more important for them to survive. Most of their losses have been CPU inventory write-downs and WSA penalties,which was a part of the agreement of the divestment of their fabs,and those fabs nearly sunk AMD as a company if they had not sold them off.

But it also means AMD still needed decent high performance process nodes - hence WSA was an agreement where AMD would guarantee a certain volume of wafers bought each year which meant GF had revenue coming to development those new nodes. Except the PC market shrunk and they had to pay additional fees so GF would keep those nodes going.

Remember,at the time TSMC was not doing CPU orientated process tech and IBM had great tech but it was not made for high volume and decent cost. It was geared towards the expensive POWER series chips.

Polaris is a modified GCN1.2 uarch GPU and evidence hints it was a 20NM design pushed onto GF 14NM. It wasn't an ideal design for 14NM.

AMD has WSA - that is why they used GF to make the GPU even though it probably was not optimal at the time - they probably needed to shift wafers to compensate for declining CPU volumes so they went mass market and cheapish just like with the ATI HD3000 series,when they made a massive redesign and brought out the HD4000 series.

Its a holding pattern set of cards.

Plus another aspect only AMD is using GF/Samsung 14NM for largish chips over 200MM2 - none of these companies would make massive commitments to continued 14NM development if AMD didn't order enough volume.

What do you think Polaris 10 and 11 are?? They are large volume parts which justifies Samsung/GF investing in 14NM and keep WSA kind of satisfied.

Edit!!

Remember - AMD is making BOTH GPUs and CPUs on Samsung/GF 14NM,which means a different set of requirements for both product lines.

[CAPTAIN_ALLCAPS]

Talk is cheap, and marketing bumf is cheaper. Furthermore, many of these features sound like they need explicit support by the game developer - and I have learned to take a dim view of technologies like that and stick to real-world performance (and the only real-world performance data we have, Doom on Vulkan, isn't terribly impressive so there had better be large amounts of untapped potential).

If AMD intended for Vega to be more Volta than Pascal they shouldn't have limited their largest Polaris chip to 232mm2, leading to 1060-level performance at best and grotesquely overpriced top end chips from nVidia for essentially an entire generation.

Trolling is cheap - your attitude in both AMD CPU and GPU threads,you seem to want to twist things as negatively as possible. Yet I can see you hyping Volta already as the next best thing - talk is cheap and we had Nvidia focus group member Rollo who started a new account before a major AMD launch, to keep pushing that Nvidia was ready to launch their latest card when eventually it didn't come for months.

Funny how just before Ryzen and Vega which will probably be launching within a few months of each other we have a new account suddenly trying to push anything AMD in the most negative way. I have been here long enough to recognise it.

But you are scared that AMD will get Vega out before Volta since your E-PEEN is threatened and you know very well if Nvidia is only launching the GTX1080TI in the next week or two,Nvidia won't be releasing a more powerful card until the end of the year,but don't worry I expect when AMD does release Vega you will make cheap talk and keep on pushing that Volta is out soon.

My attitude in AMD CPU threads? I preordered Ryzen. To replace a Phenom II X4 955. I also have an HD7950. I do not intend to replace that HD7950 yet because GP102 and GP104 are hilariously overpriced. In fact I will probably end up skipping this generation depending on how Vega 10, Vega 11, and Volta start to shape up.

Seriously, if you went to the trouble of reading my posts and then decided I was some sort of bot or troll you should really pay more attention.

I am critical of RTG because they failed to compete with anything above a 1060. Not because they couldn't, Polaris is pretty competitive per mm2 and could have at least split the 1060 and 1070 without being an out-and-out big chip or needing GDDR5X. They chose not to, and did so despite being aware that Vega could be released much later given its reliance on HBM2 and the concomitant demands of the Ryzen release. This has given nVidia the license to charge what the market can bear, not what their products are worth.

Now they are at risk of selling Vega closer to consumer Volta than consumer Pascal. That is a dangerous place to be when we aren't even certain if Vega 10 can beat GP102 (though presumably nVidia seems to think it likely given that 1080 TI is essentially a full GP102) and Volta is a total unknown.

[GrahamC]

An interesting day of GPU announcements. The price drops are welcome and the Vega info looks like it will be a card bringing new ways of working. I'll not comment on performance, that will come in the reviews.

I'm in the middle of a new Ryzen build ATM so it looks like I'll have a fine choice of new cards to pick from, an EVGA FTW style 1080Ti might look right in my build I think.

[anselhelm]

I imagine that many people, like me, struggled to watch the livestream for more than a few minutes. I love Raja, but he's not objectively a good speaker for such events where the public is watching. (For just talking to developers, I imagine he's par for the course.)

I went there hoping to see some more tidbits about Vega 11, since although Vega 10 will be a beast for 4K, I'm more interested in high-performance up to around 1440P à la Nvidia's GTX 1070, so I'm really curious to see what form Vega 11 will take.

Were it not for the fact that we're due a refresh, I'd perhaps be swayed by one of the cheaper RX 480 cards from AMD, given their price/performance ratio, but given that we're hoping to see Vega cards by May 2017, I couldn't in good faith justify buying a current generation card from AMD until the upcoming ones have launched.

Very annoying too, as my HD 6950 2GiB is now at the point where I can't play newer games in any minimum level of fidelity. Even Abzû struggles a bit in places.

[CAT-THE-FIFTH]

My attitude in AMD CPU threads? I preordered Ryzen. To replace a Phenom II X4 955. I also have an HD7950. I do not intend to replace that HD7950 yet because GP102 and GP104 are hilariously overpriced. In fact I will probably end up skipping this generation depending on how Vega 10, Vega 11, and Volta start to shape up.

Seriously, if you went to the trouble of reading my posts and then decided I was some sort of bot or troll you should really pay more attention.

I am critical of RTG because they failed to compete with anything above a 1060. Not because they couldn't, Polaris is pretty competitive per mm2 and could have at least split the 1060 and 1070 without being an out-and-out big chip or needing GDDR5X. They chose not to, and did so despite being aware that Vega could be released much later given its reliance on HBM2 and the concomitant demands of the Ryzen release. This has given nVidia the license to charge what the market can bear, not what their products are worth.

Now they are at risk of selling Vega closer to consumer Volta than consumer Pascal. That is a dangerous place to be when we aren't even certain if Vega 10 can beat GP102 (though presumably nVidia seems to think it likely given that 1080 TI is essentially a full GP102) and Volta is a total unknown.

How do you know if Polaris would scale up - you don't,and even then there is no guarantee AMD could even get a 300MM2+ GPU to work probably given the state of GF 14NM last year.

Yes,AMD is also dependent on HBM2 production schedules too.

They are also constrained by WSA too,which means that Nvidia has more freedom to choose what fabs they want. Even Nvidia used Samsung(GF licensed the tech) for only a relatively small chip months after AMD started using 14NM and GF was behind Samsung at that point.

Plus Lisa Su said Ryzen was eating up most of their R and D costs too.

Polaris a GCN1.2 design made for 20NM ported over to 14NM. It basically is a test of various power saving technologies,etc but is evolutionary over the previous generation. Look at how many respins Polaris had on 14NM - AMD were having issues producing even a 232MM2 chip on an immature process node.

Looking at the launch - they probably missed their own internal performance/watt targets due to process node issues. This is why the reference RX480 is at the limit of what the cooler can handle.

Even the RX460 is not using a full Polaris 11 chip. Even with a full high-end stack AMD marketshare fell below 20% by 2015 - they have managed to claw it back to around 30% with NO high end cards.

ATI/AMD did this before - they skipped the high-end before launching the HD4000 series which was a radical re-design of the HD2000/HD3000 series.

In the end looking at all the changes that have been discussed its a big re-design. Improved CU utilisation,reductions in VRAM usage,higher clockspeeds,move to a tiled rasterisation setup(like Nvidia has),etc which will lead to improved power consumption,etc and better performance out of less shaders.

All of these things they talk about are really ways of improving efficiency and by extension is probably what RTG have started to do what ATI did well with the HD4000/HD5000 series - Nvidia basically has done this to AMD and done very well with it.

This is the only way AMD can really compete with Pascal let alone Volta - they need to improve efficiency and also how well their CUs are used. Otherwise it will be a repeat of the Fury X vs the GTX980TI where all the theoretical processing power is sitting there not being used.

Then most enthusiasts will just get Nvidia instead.

Edit!!

Something else too - the GP104 is very efficient so Nvidia can get into laptops too. AMD with a scaled up Polaris last year probably would find it more difficult to get into such a market.

AMD is launching Vega based APUs at the end of the year to laptop first,and these apparently use Vega based IGPs,so again it makes more sense for AMD to try and just get Vega out as quickly as they could.

I would rather AMD get Vega out in a polished way - they rushed the launches of the R9 290X/R9 290 and Fury X which made them look worse than they were,and Nvidia swooped in and did well.

[Ozaron]

In the end looking at all the changes that have been discussed its a big re-design. Improved CU utilisation,reductions in VRAM usage,higher clockspeeds,move to a tiled rasterisation setup(like Nvidia has),etc which will lead to improved power consumption,etc and better performance out of less shaders.

I would rather AMD get Vega out in a polished way - they rushed the launches of the R9 290X/R9 290 and Fury X which made them look worse than they were,and Nvidia swooped in and did well.

It's time to pray for a RyZen miracle. That said, things ARE looking up. If; AMD is using some of its initial cash flow from Ryzen to fund improvements to RTG's project, RTG learned what they can from 14nm process while delivering improvements to Polaris, and maintaining focus only on their efficiency gap with NVidia, I'd say they're in for a good shout of returning to relevance vs NVidia.

Given what you've said (it does make a lot of sense) this indicates that Polaris was at its worst possible position in terms of optimization when released. And yet, it was competitive at the mainstream level, RX 480 was only 32mm bigger than 1060 and delivered marginally better performance after drivers caught up... And all of this is before the redesign, which cancels out (some of?) their power inefficiency... I'm trying not to make this sound like a pro-AMD hype post but I actually have some faith in them.

How do you know if Polaris would scale up - you don't,and even then there is no guarantee AMD could even get a 300MM2+ GPU to work probably given the state of GF 14NM last year.

ATI/AMD did this before - they skipped the high-end before launching the HD4000 series which was a radical re-design of the HD2000/HD3000 series.

I miss TeraScale.